Grade the AI closing your books, and keep the receipt.
AI-Act Ledger Attestation sits behind whatever produces your financial statements (an AI, a firm, a spreadsheet macro) and independently re-derives every statement from your general ledger, then tells you where the numbers do not tie, where the balance sheet does not balance, and exactly which line is wrong. Not a confidence score, a receipt an auditor can read.
Book an attestation pilot See a sample disagreement report
THE PROBLEM
Something produces your balance sheet and your P&L, and they look right. Most of the time they are. But when a line quietly stops tying to the ledger, when an account gets left off the page, when a subtotal rolls up wrong, or when the balance sheet does not actually balance, you tend to find out from a lender, an auditor, or a board packet, not from a dashboard. And the fashionable fix, having a second model check the first one, is just probability grading probability (a confidence number dressed up as a control), which is exactly the thing that does not survive diligence. As of August 2, 2026, the EU AI Act turns "prove what your AI did" into a board-level obligation for high-risk systems, so financials produced by an AI stop needing a nice story and start needing an independent, auditable check with a date on it.
HOW IT WORKS
Ledger Attestation does not replace your close, and it does not ask you to change your workflow, it sits behind whatever prepares the statements you already run. For each statement, it independently re-derives the numbers straight from the general ledger (deterministic, model-free at verify time) and then judges the presented statement on a fixed set of checks: does every line tie to the accounts it claims, is any account omitted or any line fabricated, do the subtotals and totals roll up, does the balance sheet balance, and does the trial balance's debits equal its credits. Where the arithmetic can decide the case, you get a real answer, agreement or a named failure. Where the source data genuinely cannot decide, it abstains and says so rather than inventing a green check. Your ledger, your close, your tools stay yours, we grade the outputs, we never take over the books.
WHAT YOU GET (the disagreement report)
For every statement where a claim exists, one auditable line: - agreement, or a business failure kind (statement does not tie, omitted account, fabricated line, arithmetic error, balance sheet unbalanced, or trial balance unbalanced), - the offending items, as a receipt with the claimed number, the re-derived number, and the accounts involved, - our own independently re-derived statement for the same ledger, so you can see the correct figures next to theirs, - the cost and runtime delta between their preparation and the deterministic re-derivation.
A report you can hand to an auditor, not a score you have to defend.
WHO IT'S FOR
Finance teams and fractional-CFO or bookkeeping platforms shipping AI into the close, first, then RegTech and audit-adjacent platforms that need a deterministic attestation layer under model-generated financials. If a wrong statement from your close can land in front of an auditor, a lender, or a board, this is for you. If a wrong number costs nothing, it is not.
PRICING (the ladder)
We price the ladder, not a single number, and every number below is a hypothesis we validate with you, not a commitment. - Pilot, a fixed-price paid proof on your own statements and ledger. You see the disagreement report and the cost delta before you commit to anything. - Value-metered, per attested statement, once the pilot proves the delta. Your close keeps running, we grade each statement. - Platform, license the deterministic attestation layer into your close or audit stack.
IP is licensed, never assigned. The verifier stays ours, the proof is yours. Any step that takes your money is gated and confirmed before it runs, nothing charges silently.
THE PROOF (dogfood)
We run this discipline on ourselves before we sell it. The re-derivation judge underneath Ledger Attestation is producer-agnostic, so it grades AYA's own generated statements the same way it grades a partner's, we eat the verifier we sell. In this build, the verify core was exercised offline across the whole taxonomy, the clean pass path plus all six failure kinds (a line that does not tie, an omitted account, a fabricated line, an arithmetic error, an unbalanced balance sheet, and an unbalanced trial balance) each fired deterministically, and a malformed ledger failed closed with an error rather than a false pass.
HONEST NOTE
Two things are true and worth saying plainly. First, the reconciliation judge is new code (there was no existing pattern that reconciled a statement to a ledger, so we wrote a minimal, fail-closed module modelled on our existing regulatory and close verifiers), and that code is marked for review and not yet wired onto a live, booted mesh. Second, it has not yet been run against a real firm's ledger export, that is the natural first live test once a read-only export is provided. Everything above the pricing is structurally built and unit-verified offline, nothing here has been billed, published, or run in production, and the pricing figures are hypotheses we intend to test with a first pilot partner.
*This page is a specification. The capability it describes is not built yet, and nothing here is a claim that it runs today.*